dsc180a
DSC180 - Auditlab
Overview
This repository contains files related to auditing opaque ML models. There is a util.py file that contains general overhead setup but most of the analysis and data collection work is completed within the notebooks. It also contains a script.py file which will act as the main querying method.
Key Links
- Code - github.com/jonathanlo411/dsc180
- Final Report - dsc180a.lojot.com/finalreport.pdf
- Website - dsc180a.lojot.com/
Usage
First, use the script.py file as a querying method. All results will be saved into a results.csv
Options:
-
-d,--debug: Writes to the log dir -
-s,--serialize: Serializes the raw HTML for ad hoc analysis
Next, you can load the results.csv as demonstrated in notebooks/analysis-template.ipynb.
Tech Stack
Web querying is handled by Selenium using a head based Chromium browser. There is an option to use the requests package in place of Selenium. Core language is in Python.
Class Details
Course Instructor: Stuart Geiger [contact]
Course Description: This group is for students interested in empirically investigating the outputs of real-world algorithmic systems for bias, discrimination, and other social issues — particularly those where the code and/or training data are not publicly available. Do facial recognition classifiers work equally well on all kinds of faces? Does a job candidate’s demographics impact which jobs they are recommended on a job search site? When you ask a generative image model to create images of a data scientist, what is the distribution by demographics?
We will study classic audits of non-algorithmic decision systems (e.g. equal opportunity hiring investigations in the 1970s) and contemporary audits of real-world ML/AI systems. We will learn various approaches to investigate such opaque systems, including auditing via synthetic training datasets, user reports, API scraping, fake/sockpuppet accounts, and headless browsers (where you programmatically control a web browser). We will also learn and discuss the legal and ethical issues around this kind of auditing, particularly around violating a platform’s terms and conditions, which are complex. All students must take and pass the UCSD/CITI IRB Human Subject Protection Training online course (Social and Behavioral Basic) by week 3 of Fall, as well as submit their proposed Winter projects to the UCSD Institutional Review Board for legal and ethical review. For a selection of readings on this topic, see a past syllabus for a related graduate course: https://auditlab.stuartgeiger.com
[1] https://dsc-capstone.org/enrollment/
[2] Course Syllabus